NIO's end-to-end smart driving large model architecture, though delayed, is almost here.

Written by | Cao Lin

Edited by | Mao Shiyang

Original content from AutoPixel (ID: autopix)

01.

Rushing Towards End-to-End Architecture

Among NIO, XPeng, and Li Auto, NIO is the only one that has not yet released an end-to-end large model architecture for smart driving. Unexpectedly, this situation is expected to change next month.

We have exclusively learned that NIO plans to release the Banyan 3.1.0 version of its intelligent system in January 2025. In this version, NIO will introduce for the first time an end-to-end large model architecture for intelligent driving.

At the end of August this year, NIO released the Banyan 3.0.0 version of its intelligent system, focusing on upgrades to the intelligent cockpit and intelligent chassis, integrating AI technology. This is currently the latest version of the Banyan system.

All eight NIO models currently on sale are developed based on the NT2.0 platform. Starting this month, NIO will gradually introduce several new models based on the NT3.0 platform, such as the ET9. The Banyan system is compatible with vehicles on the NT2.0 platform and above, while older vehicles on the NT1.0 platform cannot be upgraded for the time being and still use the Aspen system. This means that NIO vehicles produced before 2022 may not be able to use the new intelligent solution immediately.

▍Upcoming ET9

NIO's development of its end-to-end large model architecture has not been smooth sailing. A source close to NIO told us that as recently as November this year, Banyan 3.1.0 underwent a round of "major reviews" within the company. Prior to this, NIO's model team had experienced multiple delays.

The outcome of the "major reviews" was also not entirely smooth, with some issues exposed that are still being investigated. However, a source close to NIO told us that the team is currently working urgently to resolve the issues and plans to complete the release in January next year as scheduled.

However, we have learned that NIO's "parking spot to parking spot" feature may not be released along with Banyan 3.1.0. "Parking spot to parking spot" refers to seamless intelligent driving functionality from point A to point B. Currently, among domestic automakers, only Li Auto has released this feature.

A source close to NIO revealed that around early November this year, NIO was close to completing a demo version of the "point-to-point" feature internally, but there were still some issues that had not been resolved. NIO had hoped to first roll out the "point-to-point" feature to dealerships for test drives, but this did not materialize, with the functionality concentrated on specific road segments.

Compared to other peers, NIO's "point-to-point" development is indeed slightly more complex. As the only mainstream new energy vehicle maker adopting a battery-swapping model, in addition to the traditional parking spot-to-parking spot functionality, NIO also needs to develop a new version of city pilot battery swapping based on the same technology, enabling autonomous battery swapping within urban areas.

02.

Concerns About Large Models

Among the first tier of automakers, NIO has lagged behind its peers in the timing of several smart driving features.

In September this year, Huawei began rolling out ADS 3.0. On October 24, XPeng released Tianji 5.4.0, with its XNGP smart driving system fully entering the end-to-end architecture. Also at the end of October, Li Auto fully rolled out its "end-to-end + VLM" solution. A month later, at the end of November, Li Auto announced the OTA 6.5 version of its in-vehicle infotainment system, adding a parking spot-to-parking spot intelligent driving feature, making Li Auto the first automaker to achieve this functionality, ahead of Huawei and XPeng.

In terms of momentum, NIO is behind its peers. However, industry insiders have told us that the gap is not as large as outsiders might think. Strictly speaking, end-to-end refers to the process from sensor data input to control command output, without any additional steps in between. To achieve this, large models are recognized as an implementation path in the industry.

The end-to-end and large models currently implemented by domestic automakers are still in their initial stages. A smart driving practitioner told us that the end-to-end solutions released by automakers are actually composed of two or more smaller models. For example, Huawei's and XPeng's solutions each use a small model for perception and control, which is still a segmented smart driving solution in a strict sense.

Like XPeng, Huawei, and other domestic automakers, the end-to-end large model architecture that NIO will release early next year is not a strict one-stage large model. At the end of July this year, during its routine 2024 Innovation Technology Day, NIO disclosed the NIO World Model (NWM). A smart driving industry executive told us that NIO's World Model can be understood as a larger control model in terms of parameter volume.

▍NIO's NWM 'World Model'

A significant portion of the rush by automakers this year to release end-to-end and large model architectures is driven by marketing considerations. However, this will change in the coming years.

A smart driving executive from a traditional automaker previously told us that, measured by benefits, the narrow definition of end-to-end and large models is a suitable technological path for automakers to focus on, as it "enables self-iteration and leapfrog upgrades." Companies that excel in large models will create a significant gap in smart driving capabilities compared to their peers, and "this gap will continue to widen."

However, the risks for automakers in implementing this approach are also significant. The executive told us that the quality of large models heavily depends on the expertise of R&D leaders and AI experts within the system. Scientists are the soul of end-to-end and large model development. Talent that meets these requirements is not readily available to automakers, and in most cases, such talent does not align with the company's culture and DNA.

The trade-offs between risk and reward lead to different choices.

At the end of November this year, 36Kr disclosed that Li Auto had independently established an end-to-end department within the company and appointed a leader. Almost simultaneously, multiple media outlets reported that Huawei's automotive business unit had established a new technology development department, which had recruited two R&D talents through Huawei's "Genius Program" and planned to mass-produce an end-to-end large model next year.

NIO takes a conservative approach to the development of end-to-end and large models. A smart driving insider told us that as of March this year, NIO was still conducting feasibility studies on one-stage end-to-end models.

Currently, NIO's internal model team is restructured from the perception team.